Unzip into your workspace.
Use Python 3.9+ and PyTorch 2.8+. Open a terminal in the extracted package folder. Run directly; no package installation needed.
python3 -c "import torch; print(torch.__version__)"FOR YOUR NEXT PYTORCH DEBUGGING SESSION
From tensor error
to a repeatable check.
Eight broken/fixed PyTorch examples, local reports and reusable output contracts. An assembled workflow for reproducing an error, inspecting a correction and checking it again.
Work through a repair, save the report and keep a regression check beside your code. Your model stays on your machine.
Downloadable ZIP · $29 USD once · No subscription
one forward pass / a clearer trail
ILLUSTRATIVE TRACE
Flatten
CPU[2, 3, 8, 8] → [2, 192]Linear
ERRORinput: [2, 192]mat1 and mat2 shapes cannot be multipliedThe forward pass failed here. Inspect the input and the module configuration.
Real executionInspect your model, with your inputs.
Local by designNo toolkit telemetry or model uploads.
Readable evidenceKeep a report beside the code.
A SMALLER DEBUGGING LOOP
From error to
something you can inspect.
Start with the included example, then point the toolkit at a trusted local model factory.
Run an included case.
Trace the failing forward pass and write an HTML report. A reported model error is the useful result.
python3 -m tensor_preflight examples/linear.py \
--variant broken --output reports/linearThe broken example exits with code 1 and still writes its report.
Keep the evidence.
Open reports/linear.html in your browser. Run the same command with --variant fixed to compare, then preserve expected shapes as a contract.
View the actual sample report ↗WHAT'S IN THE ZIP
Runnable code.
Repeatable checks.
An executable workflow you can bring into a Python project and inspect before you rely on it.
Forward-pass tracing
A CLI and Python API to record module inputs and outputs, tensor metadata and an execution error.
HTML and JSON reports
A readable local report for inspection, with structured output for your own workflow.
Eight broken/fixed workflows
Linear dimensions, Conv2d layout, dtype, device, loss target dtype and rank, ConvTranspose geometry, and Transformer batch/sequence layout.
Reusable output contracts
Check expected output shape, dtype and device. The Transformer example uses a shape contract to catch a wrong sample count even when the forward pass returns successfully.
KNOW WHAT YOU'RE RUNNING
A trace is evidence,
not a guarantee.
This observes one execution path. Module hooks do not capture every functional operation. A successful forward pass does not prove the whole training loop is correct.
The factory loads and executes Python code. Use only code you trust. Running a model can change its state; inspect the README before tracing your own model.
Python 3.9+ and PyTorch 2.8+ required. Tested on Python 3.9.6, PyTorch 2.8.0, macOS CPU. One Linear CPU/MPS mismatch and repair also passed on this Mac. CUDA and broader accelerator compatibility remain untested.
TENSOR PREFLIGHT / DOWNLOAD
Make the next error
easier to inspect.
Local toolkit, eight debugging workflows, reports and documentation.
A ZIP you keep in your own workspace.
$29 USD / once
Secure payment through Stripe. Review the
terms and privacy details before purchase.